[Feature Request] Add Qwen3.5 architecture support

[aiindigo925]

[Feature Request] Add Qwen3.5 architecture support

Summary

Requesting support for Qwen3.5 model architecture in mlx-lm. This would enable running Qwen3.5-27B, Qwen3.5-9B, Qwen3.5-4B and their fine-tunes on Apple Silicon with MLX quantization.

Motivation

Qwen3.5 series (especially 27B and 9B) are highly capable instruction-tuned models that rival Llama-3.1 and compete well with larger closed-source models. Popular use cases:

• Agent workflows (Carnice-27B fine-tune with Hermes training)
• Coding assistants (Qwen3.5-Coder variants)
• General reasoning at efficient parameter counts

Currently these models cannot load due to qwen3_5 model type not being recognized.

Requested Models

Model	Parameters	Status
Qwen/Qwen3.5-27B	27B	Native base
Qwen/Qwen3.5-9B	9B	Native base
Qwen/Qwen3.5-4B	4B	Native base
kai-os/Carnice-27b	27B	Hermes-agent fine-tune
mlx-community/Qwen3.5-*	Various	MLX community quants

Error Reproduction

python (1/2)

from mlx_lm import load
model, tokenizer = load("Qwen/Qwen3.5-27B")

Error:

ValueError: Model type qwen3_5 not supported.

Full traceback:

File "mlx_lm/utils.py", line 71, in _get_classes
arch = importlib.import_module(f"mlx_lm.models.{model_type}")
ModuleNotFoundError: No module named 'mlx_lm.models.qwen3_5'

Environment

• Device: Mac Studio M4 Max 128GB
• OS: macOS 15.x
• mlx-lm: 0.29.1
• Python: 3.12

Related Architectures

Qwen3.5 appears to share some similarities with:

• Qwen2/Qwen2.5 (already supported)
• Llama architecture (grouped query attention, RMS norm)

Config.json structure (from kai-os/Carnice-27b):
json
{
"model_type": "qwen3_5",
"hidden_size": 6144,
"intermediate_size": 32768,
"num_hidden_layers": 36,
"num_attention_heads": 48,
"num_key_value_heads": 6,
"max_position_embeddings": 131072,
...
}

Offer to Help

We can:

• Test any PR immediately on Mac Studio M4 Max
• Provide benchmarks vs existing supported models (Llama-3.1, Qwen2.5)
• Verify tool-calling accuracy with Hermes-style prompts
• Share model weights analysis if helpful for implementation

Thank you for the excellent mlx-lm framework! 🙏

References

• Qwen3.5 blog: https://qwenlm.github.io/blog/qwen3.5/
• Carnice-27B: https://huggingface.co/kai-os/Carnice-27b
• Qwen3.5 collection: https://huggingface.co/collections/Qwen/qwen35-66b92d1f30e6ab00786f4374